arxiv:2003.09052v2 [astro-ph.im] 2 jun 2020 · were then regarded as wide- eld ccd camera mosaics...
TRANSCRIPT
Draft version June 3, 2020Typeset using LATEX default style in AASTeX61
DESIGN AND OPERATION OF THE ATLAS TRANSIENT SCIENCE SERVER
K. W. Smith,1 S. J. Smartt,1 D. R. Young,1 J. L. Tonry,2 L. Denneau,2 H. Flewelling,2 A. N. Heinze,2
H. J. Weiland,2 B. Stalder,3 A. Rest,4, 5 C. W. Stubbs,6 J. P. Anderson,7, 8 T.-W Chen,9 P. Clark,1 A. Do,2
F. Forster,8, 10 M. Fulton,1 J. Gillanders,1 O. R. McBrien,1 D. O’Neill,1 S. Srivastav,1 and D. E. Wright11
1Astrophysics Research Centre, School of Mathematics and Physics, Queen’s University Belfast, Belfast, BT7 1NN, UK2Institute for Astronomy, University of Hawaii, 2680 Woodlawn Drive, Honolulu, HI 968223Vera C. Rubin Observatory Project Office, 950 N Cherry Ave, Tucson, AZ 85719, USA4Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD 21218, USA5Department of Physics and Astronomy, Johns Hopkins University, Baltimore, MD 21218, USA6Department of Physics, Harvard University, Cambridge, MA 02138, USA7European Southern Observatory, Alonso de Crdova 3107, Casilla 19 Santiago, Chile8Millennium Institute of Astrophysics, Chile9The Oskar Klein Centre, Department of Astronomy, Stockholm University, AlbaNova, SE-10691 Stockholm, Sweden10Center for Mathematical Modeling, Universidad de Chile, Chile11School of Physics and Astronomy, 116 Church Street SE, University of Minnesota, Minneapolis, MN 55455, USA
ABSTRACT
The Asteroid Terrestrial impact Last Alert System (ATLAS) system consists of two 0.5m Schmidt telescopes with
cameras covering 29 square degrees at plate scale of 1.86 arcsec per pixel. Working in tandem, the telescopes routinely
survey the whole sky visible from Hawaii (above δ > −50◦) every two nights, exposing four times per night, typically
reaching o < 19 magnitude per exposure when the moon is illuminated and c < 19.5 magnitude per exposure in dark
skies. Construction is underway of two further units to be sited in Chile and South Africa which will result in an all-sky
daily cadence from 2021. Initially designed for detecting potentially hazardous near earth objects, the ATLAS data
enable a range of astrophysical time domain science. To extract transients from the data stream requires a computing
system to process the data, assimilate detections in time and space and associate them with known astrophysical
sources. Here we describe the hardware and software infrastructure to produce a stream of clean, real, astrophysical
transients in real time. This involves machine learning and boosted decision tree algorithms to identify extragalactic
and Galactic transients. Typically we detect 10-15 supernova candidates per night which we immediately announce
publicly. The ATLAS discoveries not only enable rapid follow-up of interesting sources but will provide completestatistical samples within the local volume of 100 Mpc. A simple comparison of the detected supernova rate within
100 Mpc, with no corrections for completeness, is already significantly higher (factor 1.5 to 2) than the current accepted
rates.
Keywords: surveys; minor planets, stars: variables: general; supernovae: general
arX
iv:2
003.
0905
2v2
[as
tro-
ph.I
M]
2 J
un 2
020
2 Smith et al.
1. INTRODUCTION
The use of ground-based wide-field telescopes equipped with large format digital cameras has opened a window
on the time varying Universe over the last decade. The late 1990s and first decade of the twentieth century saw
relatively high redshift surveys for type Ia supernovae (e.g. Schmidt et al. 1998; Astier et al. 2006) and low redshift
surveys focused on galaxy targeted searches (Leaman et al. 2011; Li et al. 2011b) The former typically employed what
were then regarded as wide-field CCD camera mosaics on 2-4m telescopes (of order 0.5◦ to 1◦ diameter field-of-view;
Perlmutter et al. 1997) with carefully selected targets for spectroscopic follow-up on the new and precious 8-10 metre
aperture facilities (Howell et al. 2005; Wood-Vasey et al. 2007).
Low redshift supernovae (SNe) were found primarily by narrow field, small telescopes (0.5 - 1. metre; Pignata et al.
2009; Leaman et al. 2011) and by well equipped and experienced amateur astronomers. These bright events were
subject to follow-up by many teams around the globe with moderate sized telescopes and many were observed and
detected from x-rays (Immler & Lewin 2003; Immler et al. 2006; Soderberg et al. 2008) to ultraviolet (Brown et al.
2009; Bufano et al. 2009) and from the mid-infrared (Kotak et al. 2006; Fabbri et al. 2011) to radio (Sutaria et al. 2003;
Ryder et al. 2004; Stockdale et al. 2007; Soderberg et al. 2010; Margutti et al. 2017). Both the low and high redshift
surveys were highly successful in their respective goals and in particular in allowing understanding of the physics of
type Ia SNe to shape their use in cosmology.
The successful techniques of both these approaches have been combined in the last decade to allow wide-field, virtually
all sky, surveys of the local Universe with 10 to 100 sq degree cameras (Quimby 2006; Law et al. 2009; Drake et al. 2009;
Chambers et al. 2016; Baltay et al. 2013; Holoien et al. 2017; Tonry et al. 2018; Bellm et al. 2019). This allows a large
volume to be searched and importantly this volume samples the local Universe, enabling extensive multi-wavelength
monitoring of bright objects. Photon starvation from distant objects tends to raise barriers to physical understanding
and development of radiative transfer simulations, shock physics models and theories of powering mechanisms (e.g.
Kromer & Sim 2009; Kasen & Bildsten 2010; Rabinak & Waxman 2011) The rate of discovery of supernovae and
transients within z . 0.1 increased dramatically and significant numbers of transients were found either in faint host
galaxies or in environments far from their high mass hosts (tens of kiloparsecs). The combination of the sheer number
of discoveries and different explosion environments probed in comparison to the historic surveys revealed an unexpected
and remarkable diversity in types of stellar explosions and merger events occurring in the local Universe. An important
scientific advance was combining the discovery of transients with rapid spectroscopic follow-up to determine redshifts,
association with host galaxies and chemical composition of the ejecta (e.g. Quimby et al. 2011; Inserra et al. 2013;
Gal-Yam et al. 2014). Furthermore, as these low redshift transients were relatively bright then multi-wavelength, time
domain follow-up has unveiled new explosion physics. And remarkably for the first time, a particle messenger other
than a photon provided an alternative alert that triggered exact identification of the source with optical-near infrared
telescopes: the joint discovery of GW170817, GRB170817A and AT2017gfo (Abbott et al. 2017a,c,b)1.
A critical piece of a wide-field imaging survey and discovery facility is the computing software and hardware that
processes the data to produce a stream of clean, real, astrophysical variables and transients. This comprises two parts,
the first being the detector detrending, source identification and astrometric and photometric calibration (e.g. Waters
et al. 2016). The principles of such processing are based on well known methods applied to narrow-field telescopes,
although wide-field camera calibration has its own challenges (Magnier et al. 2016a,c). The now ubiquitous use of
difference imaging has provided a method to identify and quantify variables and transient sources (Alard & Lupton
1998; Becker 2015; Zackay et al. 2016). However at that point the night’s catalogues of objects still requires further
processing to provide a useful information stream that can be used scientifically, and immediately, by astronomers.
This means filtering the bogus detections that inevitably exist on the difference images of a CCD, associating the
detections together and assimilating them into distinct objects and providing basic classifications (Brink et al. 2013;
Wright et al. 2015; Goldstein et al. 2015; Mahabal et al. 2019; Narayan et al. 2018). Real astrophysical detections on
a difference image comprise satellite trails, minor planets, fast moving near earth asteroids (point sources and trails),
variable and erupting stars in the Milky Way and Local Group galaxies, novae, active galactic nuclei (AGN), tidal
disruption events and extraglactic supernovae, stellar mergers and afterglows of gamma-ray-bursts. Typically most
wide-field survey and discovery projects have built a complex software system that provides a means of extracting and
classifying the sources and feeding them to spectroscopic follow-up programmes (Smartt et al. 2015; Blagorodnova
1 While multi-messenger astrophysics did begin with solar and supernova neutrinos, neither of these photon emitting objects werediscovered due to the directional information of their neutrino detections. AT2017gfo was discovered by astronomers directly following analert from a non-photonic detector.
ATLAS Transient Survey 3
et al. 2018; Graham et al. 2019). Classification of the lightcurves to filter and select objects of interest has had some
success (Lochner et al. 2016; Muthukrishna et al. 2019).
We will use the term “transient science server” for this system of software and computing hardware. The back end
of these systems which focused on the assimilation of individual detections into object lightcurves and classification
of the objects from their host object (star or galaxy) have recently been termed “brokers”. The Zwicky Transient
Facility (Bellm et al. 2019) issues a public alert stream of transient and variable sources detected on their difference
images (Patterson et al. 2019), with a view to testing methods for the upcoming Rubin Observatory Legacy Survey
of Space and Time (LSST). These alert packets are ingested and processed by a series of “brokers” to provide added
value such as context (variable star and AGN identification, host galaxy, redshift, association with known transients,
variables or outbursting source) and lightcurve classification. Currently four systems ingest these data and provide
different functionality to users, Lasair2 (Smith et al. 2019b) ANTARES3 (Narayan et al. 2018), ALeRCE4 (Forster
2019, white paper in prep.) and MARS5.
In this paper we describe how the ATLAS Transient Science Server functions and operates on a daily basis. The
ATLAS system was defined in Tonry et al. (2018) and the first major data release (a variable star catalogue) was
described in Heinze et al. (2018). ATLAS is designed, funded, and operated using NASA funds to find dangerous
asteroids which are routinely detected and submitted to the Minor Planet Center (Denneau et al. in prep). The
nature of the search for moving objects yields detections of variable and transient objects for free. To date, around 25
science papers have been published on transient and variable objects from the ATLAS data processing. This includes
the discovery of AT2018cow (Prentice et al. 2018) a candidate for a pulsational pair instability supernova (Chen et al.
2018) a possible white dwarf - neutron star merger (McBrien et al. 2019) and the lowest luminosity supernova ever
discovered (Srivastav et al. 2020). ATLAS has covered the gravitational wave skymaps of a number of LIGO Virgo
events, often covering a complete fraction of the map and discovering candidates rapidly. For example we discovered
a fast fading transient in the skymap of GW170104, which turned out to be the afterglow of a gamma ray burst
(GRB170105A) but was discovered without the high energy trigger (Stalder et al. 2017). After describing the system
in Sections 2 - 5, we briefly outline the scientific return in Section 6 and improvements for the future in Section 7.
2. THE ATLAS TELESCOPES AND SURVEY DESCRIPTION
ATLAS currently consists of two identical telescopes which are a variant of a “Wright Schmidt” design. They consist
of a 0.5 m Schmidt corrector, a 0.65 m spherical primary mirror, and a three lens field corrector sitting in front of
the filter changer and camera. They are situated at the Haleakala Observatory (HKO) on Maui and the Mauna Loa
Observatory (MLO) on the Big Island of Hawaii. The first telescope (on HKO) has been fully operational since 2015,
and the second (on MLO) was commissioned at the end of January 2017. The telescopes use two main filters in normal
survey operations. One is “cyan” (we refer to it as c) which is used during dark time and is roughly equivalent to a
composite PS1 g + r (420-650 nm). The other is “orange” (or o), used during bright time and roughly equivalent to
r + i (560-820 nm). The Haleakala unit hereafter referred to as ATLAS-HKO has both the c and o filters and the the
Mauana Loa unit (ATLAS-MLO) has only the o filter.
The cameras on each telescope (which we call an “ACAM”) each contain a STA 1600 10560×10560 pixel single CCD
detector. The pixel scale is 1.86 arcsec per pixel, and excluding the borders of the CCDs (a 30 pixel boundary) the
net field of view is 5.4◦× 5.4◦, equating to 29 square degrees. In normal survey operations, the telescope schedules are
jointly planned to tile the sky with 30 second exposures. As described in Tonry et al. (2018), the overhead from shutter
closure to beginning of the next frame is 10 sec. This is determined by the readtime of the chip (which is ∼9 sec)
since the ATLAS telescope mounts are fast and reliable. As described in Tonry et al. (2018) for telescope movements
smaller than 45◦ it takes 6.5± 0.8 sec to slew and resume tracking. Hence a total duration (expose plus overheads) of
40 sec per frame means that each telescope can take a maximum of about 1000 exposures per night during the longest
night duration of 11.5 hrs. In practice, between 800 and 1000 frames per telescope per night are routinely taken (see
Figure 1), with 900 frames covering 26 100 square degrees. This means each of the telescopes could comfortably cover
the entire visible sky from Hawaii (∼24 500 deg2: north celestial pole down to about −45 degrees declination) every
night. The entire funding for ATLAS comes from NASA to find dangerous asteroids; hence operations are designed
to optimise the detectability of moving objects. It is mandatory to collect 4 views of each field every night over a
2 https://lasair.roe.ac.uk/3 https://antares.noao.edu/4 https://alerce.online/5 https://mars.lco.global/
4 Smith et al.
span of less than 1 hour in order to reach a manageable false alarm rate. This facilitates identification of moving
objects, linkage of the tracklets, orbit estimation and submission to the Minor Planet Center. Hence the total sky
footprint covered by each telescope, 4 times each night, is around 6500 square degrees. The repeat field coverage also
enables intra night (and multi night) difference stacks to be created, allowing ATLAS to detect slow moving and static
transients at m > 20 mag across the whole sky, including the galactic plane.
The telescopes are scheduled to scan declination strips of the sky and, weather permitting, each will return to the
same footprint area every 4 nights. Hence with two units working in tandem, the whole sky visible from Hawaii is
covered every 2 nights with a 4 × 30 sec observing sequence. A typical 2 day footprint is illustrated in Figure 5 of
Tonry et al. (2018). The southern pointing limit is a boresight of δ = −47.5◦, and the southern egde of these exposures
reach δ = −50◦. The survey strategy has been defined to optimise for NEO cadence (e.g. Veres et al. 2009; Tonry
2011; Tonry et al. 2018), but a 2-day, all sky cadence is an excellent survey base for early discovery and lightcurve
classification of stationary transients.
3. ATLAS DATA PROCESSING AND OBJECT DETECTION
The data processing steps to detrend, process and to photometrically and astrometrically analyse the images are
described in Tonry et al. (2018), along with all the timings for each step (see Table 3 in that paper). Each 30 sec frame
is processed on summit computers and it takes 310 seconds to apply the first initial step of detector detrending and
photometric and astrometric calibration. It requires a further 1280 seconds to produce a matching reference image
(from the ATLAS wallpaper), subtract it from the input image, detect and classify the sources and write out a file
containing the detections with S/N > 5σ. Hence a total of 27 minutes is required from the time of shutter opening
for an exposure to the final processed file containing a list of detections. The processing runs multi-threaded, and not
in serial, hence 27 minutes is the typical delay time from observing to initially quantifying detections on the cameras.
We have the capability of reducing the elapsed time by factor of ∼2 through deployment of multi-threading on the
summit computers, but currently do not do this since the primary science goal of asteroid detection has a latency of
∼30 mins due to the 4 visits.
A key part of transient object detection is subtraction of a reference sky frame from the input image to allow for
rapid and automated detection of variables, moving objects and transients (Tomaney & Crotts 1996; Alard & Lupton
1998). In ATLAS we refer to this reference sky as the ATLAS wallpaper, and we currently rebuild this wallpaper
about once per year. We store the wallpaper in files of square degree size, and routines allow any ATLAS pointing
on the sky to trigger a reconstruction of a wallpaper reference image of the appropriate size and alignment. We use
a modified version of the image subtraction algorithm HOTPANTS (Becker 2015) to subtract the matched, convolved
reference image from the target frame.
The ATLAS reduction pipeline has a custom built point-spread-function fitting routine that runs on the difference
images to produce flux measurements of all sources that are detected at 5σ or more above the background noise.
This routine (written in C) is called tphot, and is based on the algorithms discussed in Tonry (2011) and Sonnett
et al. (2013). This produces one headed ASCII table per image (internally appended with .ddc), and the contents are
summarised in Table 1. Some of these quantities are used in an initial filtering process to remove the bogus sources on
the difference images (see Section 4.3).
On a typical night, a single ATLAS telescope produces about 900 catalogue files (one per exposure). Each catalogue
file contains on average about 20 000 detections, so about 10-20 million detections are ingested per night from the two
telescopes combined. The vast majority of these sources are not astrophysically real objects and a series of quality
control procedures and filters, together with machine learning algorithms are run to reject the false positives and
leave real transient objects. Figure 1 illustrates the number of detections ingested from these .ddc files per night and
example breakdowns of their classifications are in Table 2.
4. THE ATLAS TRANSIENT SCIENCE SERVER
4.1. Object definition and spatial indexing
The essential functions of a hardware and software system to deliver real transients from a synoptic wide-field survey
such as ATLAS includes amalgamating detections into objects, removing false positives, assimilating lightcurves, and
associating the newly discovered objects with known catalogues. We have built the ATLAS Transient Science Server
to provide all these functions. The Server is currently based on individual detections from single ATLAS exposures,
ATLAS Transient Survey 5
Table 1. The contents of a .ddc file for a difference image contains all detections (5σ significance or greater), each with thefollowing measured attributes
.ddc file entry meaning
RA astrometrically calibrated RA in decimal degrees of centroid
Dec astrometrically calibrated DEC in decimal degrees of centroid
mag AB magnitude from PSF fit
dmag error in AB mag
x x pixel on CCD of centroid
y y pixel on CCD of centroid
major PSF FWHM major axis [pixels]
minor PSF FWHM minor axis [pixels]
phi angle of major axis, counter-clockwise from x axis [degrees]
det integer indicating fixed, free, trailed, streak, or negative flux analysis
chi/N reduced χ2 of the PSF fit
pvr vartest probability of source being a variable star
ptr vartest probability of source being a moving object or astrophysical transient
pno vartest probability of source being just a noise fluctuation
pbn vartest probability of source being a vertically extended electronic “burn” artifact
pcr vartest probability of source being a cosmic ray hit
pxt vartest probability of source being an electronic crosstalk artifact
psc vartest probability of source being a star subtraction residual “star scar”
pmv vartest probability of source being a moving object (subset of ptr)
pkn vartest probability of source being a known minor planet
dup integer indicating whether this PSF measurement should be kept or culled
WPflx forced flux measured in the wallpaper (reference sky) at the position of the transient
dflx error in WPflx
but in future we intend to combine the information from the nightly difference stacks to allow us to detect transient
objects earlier.
The ASCII catalogue files of objects detected on the difference images (the .ddc files described in Section 2) are
transferred to a machine at Queen’s University Belfast (QUB). This is a 28 core server with 128GB RAM (also used
as the ingestion machine). The files are transferred (using parallel download threads) in the background every 30
minutes between 04:00 UT and 22:00 UT (corresponding to the earliest start of a Hawaiian observing night at 18:00
and cleaning up after the night is done at 12:00 next day). Just before an ingest cycle begins, an extra check is done
to make sure we have the very latest dataset. The reduced, calibrated images and their associated subtracted frames
are also transferred (also in parallel) to QUB during the Hawaiian night. Typical end to end data rates from the
University of Hawaii Manoa campus to the QUB machine are 40 MBytes/sec. Since individual download threads are
throttled by the network infrastructure, using 8 parallel threads ensures we use the maximum available bandwidth.
Each of the 8 threads achieves 5 MBytes/sec and using more than 8 threads brings very little gain in download speed.
The .ddc file contains a header comprising 37 key-value pairs describing metadata of the exposure, such as the tele-
scope pointing coordinates, exposure MJD, exposure time, filter, PSF FWHM, zero point and 5σ limiting magnitude.
This is followed by a headed table with each entry corresponding to a single detection (see Table 1). The metadata
and detections are automatically read by a C++ ingester (generated by python code which reads the .ddc structure) at
QUB which connects to a local MySQL database. Detections are first checked via cone searching (see below) against
previously ingested objects. If a detection is within 3.′′6 (twice the ATLAS pixel size) of an existing object, it is ingested
6 Smith et al.
58000 58400 588000
200
400
600
800
1000N
umbe
r of
Exp
osur
esMLOA
58000 58400 588000
1
2
3
4
5
Det
ectio
ns In
gest
ed×
107
MLOC
58000 58400 58800MJD
0
1
2
3
4
5
Mea
n de
tect
ions
per
expo
sure
×10
4
MLOE
2017 2018 2019 2020Date
57600 58000 58400 588000
200
400
600
800
1000
Num
ber
ofE
xpos
ures
HKOB
57600 58000 58400 588000
1
2
3
4
5
Det
ectio
ns In
gest
ed×
107
HKOD
57600 58000 58400 58800MJD
0
1
2
3
4
5
Mea
n de
tect
ions
per
expo
sure
×10
4
HKOF
2016 2017 2018 2019 2020Date
Figure 1. A, C, E: Number of exposures, number of detections ingested and average number of detections per exposure forMLO. The detections are the number of sources at 5σ or greater that are present on the difference images (in the .ddc files). B,D, F: The same for HKO. The gap in the HKO data was caused by an ice storm on Haleakala. See also Figure 7. The rise in theaverage number of objects per exposure around MJD=58000 is because of a change of detection file schema and improvementsin source extraction.
and formally associated with that object, becoming another point on the lightcurve. If there are no objects within
the specified proximity of the ingested detection, a new object is created and the detection is associated with the new
object. To speed up the ingest, multiple ingester processes are run in parallel using a python wrapper (to utilise all
the cores of the ingestion machine).
Hence all records in the .ddc files are preserved in the database. The critical values ingested are RA, Dec, calculated
HTM ID (see below), mag and dmag. Other detection properties are also ingested to facilitate post ingest cutting.
The detection data are spatially indexed using Hierarchical Triangular Mesh (Szalay et al. 2005). Since the vast
majority of queries are cone searches of the order of a few arcsec in size, HTM Level 16 (2 × 4(16+1) triangles, hence
mean triangle area = 15.6 square arcsec) is used. The HTM value, calculated for every detection’s RA and Dec upon
ATLAS Transient Survey 7
0 100 200 300 400 500 600 700 800 900 1000p
100
101
102
103
104
105
106
Freq
uenc
yA pvr
ptr
0 100 200 300 400 500 600 700 800 900 1000p
100
101
102
103
104
105
106
Freq
uenc
y
B pscpbnpxt
0 100 200 300 400 500 600 700 800 900 1000p
100
101
102
103
104
105
106
Freq
uenc
y
C pnopcr
0 100 200 300 400 500 600 700 800 900 1000p
100
101
102
103
104
105
106
Freq
uenc
y
D pmvpkn
Figure 2. Histograms of the vartest detection values for night of MJD=58816 (2019 November 29). The output of vartest
produces values for each parameter 0 < p < 999. Of the 31 million detections that night, only those that form objects comprisingat least three distinct intra night detections (2.4 million detections, 602 000 objects) are shown. A: Histogram of the valuesof pvr and ptr. The dotted line is the current threshold above which a detection is regarded as good quality for the purposeof detecting variable objects and transients. B: Histogram of the values of psc, pbn and pxt. C: Histogram of the values ofpno and pcr. D: The values of pkn and pmv are treated separately and the histogram of values is plotted. The value of pkn isnormally 999 or 0 (i.e. a detection is associated with a known object or it is not), whereas the value of pmv will span the entirerange. Of 160 000 known minor planet detections that night, 31 000 detections were of movers slow enough to be aggregatedinto 4 800 objects comprising 3 or more detections.
ingest, is an integer and this column is indexed inside the relational database. The C++ HTM library provides a cone
searching API which returns the ranges of triangles that overlap a cone on the sky. This is used to create an SQL
WHERE clause. All objects in the database that have those triangle IDs are returned to the requesting code. The cone
search code then just needs to eliminate objects outside the requested radius (i.e. the edges of the resultant jagged
cone are smoothed).
4.2. Selection criteria, machine learning and classification
After ingest, selection criteria and algorithms are applied to select real astrophysical transients from the detection
stream.
Multiple detections: since the standard quad set of 30 sec images are separated across a 1 hr window of observation,
spatial coincidence of the sources is used to reject some moving objects and other non-astrophysical contaminants.
Three or more good quality, co-spatial detections are required to define an object and all three must be on the same
night. They must also not be within 100 pixels of the edge of the detector. There is no selection applied to the
coherence of the magnitudes of three detections, and we do find rapidly evolving transients (over the course of 1 hour),
all of which we have identified as M-dwarf flares apart from one GRB afterglow (Stalder et al. 2017).
8 Smith et al.
The main benefit of reducing the triggering threshold to 2 detections rather than 3 would be that we may discover
and detect objects earlier. Given the 2 day cadence of ATLAS, objects with only 2 detections on a single night and no
other 5σ detections are either real transients that just rise to the sensitivity limit in that window, earlier detections
of objects that continue to rise (and are subsequently detected with ≥ 3 detections afterwards) or spurious artefacts.
We carefully checked 50 real objects from 2020 April, and found that 40% could potentially have been discovered
earlier had we used a 2 detection minimum criterion rather than 3. The mean time difference was a gain of 4 days.
However while that means 40% of objects have the potential to have been discovered earlier, it does not mean they
would definitely have been promoted. The early detections would need both a high RB score (see below) and to be
confidently judged by human scanners as real. Our human scanners will often wait for an extra night of data if the
images are ambiguous. This gives a sense of the maximum gain if we were to go to a criterion of 2 images. It is
possible, but the penalty is a large increase in false positives. For example, the figures shown in brackets in Table 2
illustrate that human scanners would have been presented with over 24 times more objects. Seven objects submitted
to the IAU TNS by other surveys (PS1 and ZTF), all fainter than 19 mag at peak, would have been triggered with a
2 detection threshold by ATLAS, all with a high RB factor, but after the other surveys had already reported them.
Source classification: A pixel-based classification of candidate source detections in the ATLAS difference images
is performed using a program called vartest, written in C (Heinze et al. in prep), which runs in Hawaii after the
difference image detections have been extracted. vartest attempts to classify sources into several categories of real
and spurious objects. It is not a machine learning code, but rather uses a logical cascade of manually tuned thresholds
and analytical probability estimators applied to various pixel-based metrics. These metrics include centroid offset and
flux ratio between the original and difference images; elongation of the PSF; pixel-coordinate association with bright
stars; ratio of nearby bright and dark pixels; and many others. The classification is done probabilistically, with p = 0
being the lowest probability and p = 999 being the highest. The categories of real objects are known asteroids (pkn),
unknown asteroids or transients (ptr), and variable stars (pvr). Objects with nonzero values of ptr are also assigned a
probability of being a fast-moving (i.e. trailed) detection (pmv). The categories of spurious detections are noise (pno),
electronic column artifact or “burn” (pbn), cosmic ray (pcr), electronic crosstalk (pxt), and star subtraction residual
or “star scar” (psc). These are summarised in Table 1. The sum of the values for a single detection (excluding pmv
and pkn, which are subsets of ptr) will not exceed 999. vartest also identifies duplicate detections in its input file,
and attempts to indicate the “best” detection (stored in the .ddc dup column).
From the vartest set of parameters we define potential “good quality” detections as those with values pvr > 50 or
ptr > 50. If either of these two criteria are not met, the detection does not count toward the 3 minimum good quality
detections within a night required to define an object. Additionally, the detection must NOT belong to a known mover
(pkn < 500) or a possible mover (pmv < 500) or be marked as crosstalk (pxt < 500). The value of det must not
indicate negative flux and the value of dup must not indicate a deprecated duplicate value. Figure 2 shows histograms
of the vector values for the full (dark) night of MJD=58816 (2019 November 29). The data rows used in the plots are
only those which comprised objects of 3 or more distinct detections. The dotted line on the ptr and pvr plot shows
where the quality cut is currently set.
Object recentering: If an object passes the above criteria, the detection closest to the mean coordinates of the object
is identified as the representative detection for that object. This ensures that subsequent detections are associated
with the correct objects. (The code is periodically rerun to make sure the object is properly centred.)
Ephemeris checker: An initial cross-match with known minor planets is done in Hawaii (within vartest) using a
local copy of the Minor Planet Center (MPC) orbital elements and a 9.3 arcsecond (5 pixel) association radius. This
produces an entry in the .ddc, labelled as pkn which is the probability that the detection is that of a known asteroid.
However, this occasionally misses some asteroids, and does not currently label any comets and thus we implement a
further check. Three times weekly, at QUB, the Minor Planet Center (MPC) orbital element database is downloaded
and converted into an XEphem database file. At the same time, the comet elements, provided by the MPC in XEphem
format already, are appended to the converted downloaded file. After each batch of data is ingested, the latest detection
of an object that passes previous cuts is passed to code which locally generates the solar system object ephemeris from
the orbital element data based on the MJD of the exposure. The detections are grouped by MJD to minimise the
number of times an ephemeris needs to be generated. A cone search is then run against the ephemeris to identify
any solar system object (asteroid or comet) that may lie within a specified search radius (currently 10 arcsec). If
it lies within the search radius, the object is tagged as a probable known mover and labelled with the solar system
object ID. The ephemeris check is done offline without interrogating the MPC servers (apart from the periodic data
ATLAS Transient Survey 9
downloads) to minimise MPC load. Additionally, moons of the giant planets are not currently recorded in the MPC
database of orbital elements and therefore fail to be found in this basic check. For this reason, the positions of all the
satellites of the giant planets have been generated6 for every hour for the next 10 years and these are stored in a local
database table at QUB. A simple cone search is run against this for the most recent detection of each object, with an
additional temporal check of any matched object. If a matched planetary satellite is within 20 arcsec and half a day
of the timestamp of the detection, the object is labelled as a probable mover.
Although every effort is made to reject moving objects from the flow of data to human scanners, two serendipitous
solar system discoveries have been made with the ATLAS Transient Science Server. We discovered unusual activity of
the asteroid (6478) Gault (Smith et al. 2019a; Kleyna et al. 2019), arising from a comet-like “tail” which provided a
locus of detections offset from the asteroid itself. We also discovered a distant previously unknown long period comet
C/2019 K7 (Smith) (Heinze 2019a,b), at 4.5AU perihelion which was therefore slow moving and made its way through
the Transient Science Server as a possible stationary transient.
Sherlock: Once an object is defined, a boosted decision tree algorithm (internally known as Sherlock) mines a library
of historical and on-going astronomical survey data and attempts to predict the nature of the object based on the
resulting crossmatched associations found. One of the main purposes of this is to identify variable stars, since they
make up about 50% of the objects (see Table 2), and to associate candidate extragalactic sources with potential host
galaxies. The full details of this general purpose algorithm and its implementation will be presented in an upcoming
paper (Young et al. in prep), and we give an outline of the algorithm as implemented within the ATLAS Transient
Science Server here.
The library of catalogues contains datasets from many all-sky surveys such as the major Gaia DR1 and DR2 (Gaia
Collaboration et al. 2016, 2018), the Pan-STARRS1 Science Consortium surveys (Chambers et al. 2016; Magnier
et al. 2016a,c,b; Flewelling et al. 2016) and the catalogue of probabilistic classifications of unresolved point sources by
Tachibana & Miller (2018) which is based on the Pan-STARRS1 survey data. The SDSS DR12 PhotoObjAll Table,
SDSS DR12 SpecObjAll Table (Alam et al. 2015) usefully contains both reliable star-galaxy separation and photometric
redshifts which are useful in transient source classification. Extensive catalogues with lesser spatial resolution or colour
information that we use are the GSC v2.3 (Lasker et al. 2008) and 2MASS catalogues (Skrutskie et al. 2006). Sherlock
employs many smaller source-specific catalogues such as Million Quasars Catalog v5.2 (Flesch 2019), Veron-Cett AGN
Catalogue v13 (Veron-Cetty & Veron 2010), Downes Catalog of CVs (Downes et al. 2001), Ritter Cataclysmic Binaries
Catalog v7.21 (Ritter & Kolb 2003). For spectroscopic redshifts we use the GLADE Galaxy Catalogue v2.3 (Dalya
et al. 2018) and the NED-D Galaxy Catalogue v13.17. Sherlock also has the ability to remotely query the NASA/IPAC
Extragalactic Database, caching results locally to speed up future searches targeting the same region of sky, and in
this way we have built up an almost complete local copy of the NED catalogue. More catalogues are continually being
added to the library as they are published and become publicly available.
At a base-level of matching Sherlock distinguishes between transient objects synonymous with (the same as, or very
closely linked, to) and those it deems as merely associated with the catalogued source. The resulting classifications are
tagged as synonyms and associations, with synonyms providing intrinsically more secure transient nature predictions
than associations. For example, an object arising from a variable star flux variation would be labeled as synonymous
with its host star since it would be astrometrically coincident (assuming no proper motion) with the catalogued source.
Whereas an extragalactic supernova would typically be associated with its host galaxy - offset from the core, but close
enough to be physically associated. Depending on the underpinning characteristics of the source, there are 7 types of
predicted-nature classifications that Sherlock will assign to a transient:
1. Variable Star (VS) if the transient lies within the synonym radius of a catalogued point-source,
2. Cataclysmic Variable (CV) if the transient lies within the synonym radius of a catalogued CV,
3. Bright Star (BS) if the transient is not matched against the synonym radius of a star but is associated within
the magnitude-dependent association radius,
4. Active Galactic Nucleus (AGN) if the transient falls within the synonym radius of catalogued AGN or QSO.
6 via Giorgini, JD and JPL Solar System Dynamics Group, NASA/JPL Horizons On-Line Ephemeris System, http://ssd.jpl.nasa.gov/?horizons, data retrieved 2019-02-21.
7 https://ned.ipac.caltech.edu/Library/Distances/
10 Smith et al.
5. Nuclear Transient (NT) if the transient falls within the synonym radius of the core of a resolved galaxy,
6. Supernova (SN) if the transient is not classified as an NT but is found within the magnitude-, morphology- or
distance-dependant association radius of a galaxy, or
7. Orphan if the transient fails to be matched against any catalogued source.
For ATLAS the synonym radius is set at 1.5′′. This is the crossmatch-radius used to assign predictions of VS,
CV, AGN and NT. The process of attempting to associate a transient with a catalogued galaxy is relatively nuanced
compared with other crossmatches as there are often a variety of data assigned to the galaxy that help to greater inform
the decision to associate the transient with the galaxy or not. The location of the core of each galaxy is recorded so
we will always be able to calculate the angular separation between the transient and the galaxy. However we may also
have measurements of the galaxy morphology including the angular size of its semi-major axis. For ATLAS we reject
associations if a transient is separated more than 2.4 times the semi-major axis from the galaxy, if the semi-major axis
measurement is available for a galaxy. We may also have a distance measurement or redshift for the galaxy enabling
us to convert angular separations between transients and galaxies to (projected) physical-distance separations. If a
transient is found more than 50 Kpc from a galaxy core the association is rejected.
Once each transient has a set of independently crossmatched synonyms and associations, we need to self-crossmatch
these and select the most likely classification. The details of this will be presented in a future paper (Young et al.
in prep). Finally the last step is to calculate some value added parameters for the transients, such as absolute peak
magnitude if a distance can be assigned from a matched catalogued source, and the predicted nature of each transient
is presented to the user along with the lightcurve and other information (see Figure 5).
We have constructed a multi-billion row database which contains all these catalogues. It currently consumes about
4.5TB and sits on a separate, similarly specified machine to that of the ATLAS database. It will grow significantly as
new catalogues are added (e.g. Pan-STARRS 3π DR2, VST and VISTA surveys, future Gaia releases etc).
Cross-matching with other surveys: The flagged objects are also cross-matched against a database of transient objects
discovered by other surveys and tagged if they are already known. We compile a local database from the IAU TNS
and also cross-match against objects from ZTF (Bellm et al. 2019), which are contained in the Lasair broker and
are likely to be supernova or extragalactic transients of some sort (Smith et al. 2019b). Hence users know almost
instantaneously if an object has been discovered by another survey and registered in the TNS as a discovery, or if it
has multiple detections in the ZTF public stream. This crossmatch runs at the end of every ingestion cycle (at least
4 times a day) with a match radius of 3.′′0.
We also trawl “The Astronomer’s Telegram”8 text reports (ATels) for information connected to ATLAS transients.
There are two distinct searches performed against the human-written, plain-text content found in the title and body
of each ATel; a name-search and a coordinate-search. As ATLAS and all other major transient search projects have
consistent transient naming schemes, we have found it possible to compile a set of regular-expressions that match
most of the transient names reported in the ATels. All matched names are recorded in a separate table to the
original database table hosting the ATel titles and content, with ATel number as the relational key between tables.
Furthermore a long and complex regular expression is employed to extract out all sets of equatorial coordinates found
within ATels. This expression has to accommodate for the plethora of formats of sexagesimal coordinates and also
for the surprisingly common human-errors found in those reports. The coordinates are homogenised into decimal
degree format and recorded in another database table which is then indexed using the familiar HTM indexing scheme.
Whenever a new ATel is released, transient names and coordinates can then be matched against ATLAS transient
names, pseudonyms from other surveys and coordinates. The discovery of a new ATLAS transient also triggers a name
and coordinate-search for matches in past ATels. The resulting ATel matches are reported on the ATLAS transient
pages providing direct web-links to the associated ATel reports. However use of “The Astronomer’s Telegram” for
registering, announcing and classifying extragalactic transients and supernovae is dwindling, being replaced by a more
robust and accessible database system in the IAU TNS. The latter allows full photometric and spectroscopic registration
in an automated way, combined with a well maintained database and links to WISeREP (Yaron & Gal-Yam 2012). It
also allows reports to be written and released, as AstroNotes (Gal-Yam et al. 2019), all of which are indexed on the
8 http://www.astronomerstelegram.org
ATLAS Transient Survey 11
NASA Astrophysics Data System and are citable, hence providing all the functionality of ATels but with added data
curation and searching value.
Machine learning image recognition: For the objects that remain after the stars are removed, we construct a triplet
of “postage stamps”. The target image, reference and difference images centred on the detections are produced of size
6.2 arcmin (200×200 pixels). A machine learning algorithm, trained to recognise “good” subtractions (i.e. PSF-like
objects in the centre of the difference stamp) is run on the multiple (at least three) difference images associated with
each object. A median score between 0 (bogus) and 1 (real) is applied to each object. All objects with a score
above a chosen threshold are passed to human scanners. The training sets were built from the 20×20 pixel core of
difference image stamps centred on 200 000 known asteroids (real) from data sampled over six separate nights and
a random selection of the 20×20 cores of difference image stamps of 600 000 objects that were not known variables
(bogus). Two training sets were built, one for each telescope, both of which have similar performance, though the
HKO dataset was built using three cyan and three orange nights, whereas the MLO classifier only uses orange nights
(because it only has the one filter). Figure 3 shows the performance of the classifier, displaying a plot similar to a
receiver operator characteristic (ROC) curve and a hypothesis distribution. We began using the machine learning
algorithm as described in Wright et al. (2017) and Wright (2015) (random forest and sparse filtered neural network)
but now use a Keras (Chollet et al. 2015) convolutional neural network with a Tensorflow (Abadi et al. 2015) backend.
Convolutional neural networks (CNNs) were applied successfully for transient discovery and detection in image data by
Cabrera-Vives et al. (2017), showing excellent promise for use in high data volume surveys (see also Reyes et al. 2018).
Our Keras/Tensorflow CNN, which is a modification of an example model used to recognise the MNIST database of
handwriting digits, consists of three 2D convolutional layers with 16, 32 and 64 filters and a kernel size of 2×2 pixels.
Each convolutional layer is followed by a maxpooling layer that pools over 2×2 pixel regions. The final pooling layer
feeds into a dropout layer (to prevent overfitting) with a dropout rate of 0.3, which is followed by a Dense (fully
connected) layer with 500 units and another dropout layer with a dropout rate of 0.4. The output layer is a Softmax
classifier that produces two outputs; the probability the detection is real and the probability the detection is bogus.
All units other than those in the output layer are rectified linear units (relu). “Same” padding (Dumoulin & Visin
2016) is used throughout and the model is trained with the categorical cross entropy loss (Hastie et al. 2009) and
optimised with adam (Kingma & Ba 2014).
We are of course free to set the threshold for the RB factor to any value to optimise either completeness or purity.
In reality we run our daily transient searches by selecting objects with RB ≥ 0.2, in the knowledge that this results
in 96% completeness. As can be seen in the example numbers in Table 2, this reduces the objects that humans will
scan by a factor 20, from about 9000 to 300, leaving quite a manageable number per day. The classifier is periodically
retrained as the training sets are refined. This is necessary because of changing conditions (e.g. PSF improvements
over time). Selection of representative “bogus” data is also improving as human scanners log more examples where
machine classification has failed. We foresee improvements in the training sample by using all the images of all objects
we now consider as real stationary transients. This may improve the performance of the classifier on highly varying
spatial backgrounds (which leave noise residuals in the difference images). With close to 8000 real objects in our
database and at least 20 images per object this will provide a rich and useful training set. Employing an unsupervised
method could allow clusters of real and bogus objects to emerge and further improve performance.
Forced photometry: For each potential extragalactic transient, or orphan transient, that has minimum of 3 detections
at 5σ significance, and a machine learning score above our chosen threshold (i.e. has passed the above filters), we
carry out forced photometry on historic images up to 30 days before initial discovery at the position of the transient
(using the rolling average of its RA and DEC in the unforced .ddc files). This normally results in one or more pre-
discovery epochs with multiple 2-4σ detections on the individual 30 sec frames. We calibrate and measure the forced
photometry in flux (µJansky), since conversion between AB magnitudes is trivial, allowing low significance positive
values, zero values, and negative values to provide real meaning (which is lost when converting to Pogson magnitudes).
This is invaluable in constraining the explosion epoch for nearby objects and providing constraining early data for all
transients. This information is presented to the user along with all the above information, as explained in the next
section (see Figure 5 for an example). The forced photometry is again PSF based, with tphot used in the same way as
for unforced PSF flux measurements, but simply with the centroid position fixed and the PSF modeled from the input
image before subtraction. Forced photometry is also done against a stack of the four intra-night difference images for
an object, which can help to tie down the discovery epoch for an object.
12 Smith et al.
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0Realbogus Factor
102
103
104
105
Num
ber
of O
bjec
tsA bogus
real
0.000 0.025 0.050 0.075 0.100 0.125 0.150 0.175 0.200Missed detection rate
0.000
0.025
0.050
0.075
0.100
0.125
0.150
0.175
0.200
Fals
e po
sitiv
e ra
te
B
Figure 3. A: Histogram showing the classifier performance against the training set (for the Haleakala (HKO) classifier). Thedotted line shows the RB factor threshold which corresponds to a 4% missed detection rate of real objects. The real/bogusvalue threshold is RB = 0.2, below which objects are regarded as bogus. (Note the logarithmic vertical scale.) B: Detectionvs Error tradeoff. The dotted line shows the false positive rate (3.65%) at the set missed detection rate of 4%, based on a testsubset of the training set. This is for an RB = 0.2 threshold for real/bogus selection.
Table 2. Example nightly ingest numbers for two ATLAS telescopes on the night of MJD=58816 (2019 November 29). Goodquality detections are defined as having pvr > 50 or ptr > 50 and not a known or possible mover, must be positive flux andnot crosstalk. The detection must also not be within 100 pixels of the edge of the detector. Note that the number of “good”objects for this night is higher than average. This was possibly because of dark time, the slightly longer winter night, and a fullnight of good quality difference images. The figures assuming 2 spatially coincident detections rather than 3 are also shown inbrackets.
Stage Number (2 detection Number)
Detections on the difference images (≥ 5σ in the .ddc files) 3.1× 107
Single detection objects (probably noise, edge, streaks and 1.6×105 known movers)
2.1× 107
Objects with 3 (2) or more spatially coincident detections 6.2× 105 (2.6× 106)
Objects with 3 (2) or more spatially coincident, good qualitydetections
2.2× 104 (3.6× 105)
Objects remaining after stars removed 9 450 (68 542)
Objects remaining after AGN removed 9 300 (68 068)
Objects remaining after untagged known slow movers removed(e.g. comets)
9 280 (68 008)
Objects remaining after ML and RB ≥ 0.2 applied 340 (8 341)
Good objects after human scanning 40
4.3. User interaction and object selection
The user is presented (via the web interface) with a triplet of postage stamps: the target image (typically 4 per
night), the reference image used as the template for subtraction, and the difference image. The scanners inspect these
and use a web form to decide whether or not they are indeed real or bogus transients that have a high machine learning
score. Figure 4 shows examples of the triplet of images presented to the users. As well as these triplet images, the
user is presented with a range of other visualisation plots and tools which include the forced photometry, Aladin-lite
(Bonnarel et al. 2000; Boch & Fernique 2014) plugin showing the Pan-STARRS1 3π archive image (if above δ > −30◦,
otherwise it defaults to the DSS), Galactic coordinates, a summary of the Sherlock cross-matching results (along with
host galaxy information). Figure 5 shows examples of the panels displayed to users. At this point, human scanners
decide whether to log the object in the ATLAS “Good Objects” list, reserve judgement and save it to a holding list
ATLAS Transient Survey 13
(called “Possible Objects” ) to wait for further data, or to archive it. The object remains in the database even if
archived, and can be retrieved. The Transient Name Server (TNS) python API9 is used to submit objects promoted
to the TNS as soon as objects have been indentified as “Good” by the scanners. All future detections from ATLAS
that are spatially associated with an object get added to the lightcurve, and forced photometry runs on all objects
in the “Good Objects” and “Possible Objects” list once a day, and the lightcurve plots are updated. A summary of
nightly ingests are given in Table 2.
GW Event Tagging: Software has been written (in python) to download the latest LIGO-Virgo GW event Healpix
maps and calculate which ATLAS objects lie within the 90% contour region of a GW event. Objects are tagged with
the event ID, event MJD and contour value (in 10% divisions) if they are within -10 to +21 days of the GW event. E.g.
“Possible GW Events Association: GW190425 (MJD 58598.3458914 - 60% contour).” The ATLAS web interface has
an additional filter facility that can be used to search for named GW events and display lists of objects that lie within
the 90% region. This facility can be used to prioritise the scanning of relevant objects after a GW event. This was
used to discover the afterglow of GRB 170105A which was in the skymap of the binary black hole merger GW170104
(Stalder et al. 2017).
4.4. Hardware
The processing of data is done by two rack servers - a 28 core server with 128GB RAM for the ATLAS data
processing and another identical but separate machine to host the database. The servers are linux based (CentOS7)
and are equipped with 26TB (RAID6) of spinning disk (for postage stamp storage) and 16TB of NVME SSD in a
RAID0 configuration (for database storage, to reduce I/O latency). The database SSDs are backed up to spinning
disk twice a week, so loss of an SSD, while destructive to the array, would not result in permanent loss of data.
ATLAS postage stamps are stored on a 26TB RAID6 spinning disk array, local to one of the above servers, while the
full sized reduced ATLAS images (input, reference and difference) are stored on a bank of 4 RAID6 70TB dedicated
network attached storage arrays.
The webserver requires relatively little CPU or memory, so is a separate virtual machine (running on shared university
hardware) with an allocation of 4 cores and 8GB RAM. Local webserver data is stored on a small SSD with postage
stamp images accessed via NFS from the above 26TB array.
Raw ATLAS data (currently about 700TB) is backed up directly from Hawaii to the QUB petabyte scale HPC
cluster. Periodic ATLAS database dumps and postage stamps are also stored there.
4.5. Software
Most of the software is written in Python, with C++ (ingester) and javascript (dynamic lightcurve rendering) com-
ponents. The ingester, post ingest cutting, context classification and external survey crossmatching are marshalled by
bash scripts called via CRON.
Ingester: The ingester is written in C++ (with the template functions generated by python code) and dependent on
the C++ HTM library. The code is spawned using a Python Multiprocessing wrapper.
Post Ingest Cutting and Crossmatching: The post ingest processing, context classification (Sherlock) and external
crossmatching code (including minor planet rejection) is written in python, with extensive utilisation of Astropy
(Astropy Collaboration et al. 2013; Price-Whelan et al. 2018). Most of the code is called via the Python Multiprocessing
module to make maximum use of our CPUs.
Database: The database technology is MySQL. The “back-end” storage engine chosen for most tables is the non-
transactional MyISAM engine. Although this is becoming obsolete in future, with users strongly being advised to
move to InnoDB (transactional storage), the MyISAM engine is still the preferred solution for ATLAS for the time
being. Its main advantage is that the database files are a third of the size of InnoDB database files. Given that the
database is currently 6TB in size, we would require 18TB of SSD storage to run the InnoDB version just to stand
still. Another advantage is that the database files can be copied directly from one machine to another without having
to apply transactional logs (assuming the source and destination machines are running the same version of MySQL).
Additionally, the way the database is currently designed (with primary keys of objects represented by their location
in the sky) does not lend itself well to the InnoDB model of database row insertion (though this would be relatively
easy, albeit time consuming, to rectify by the creation of a new running counter primary key). One final advantage
9 https://wis-tns.weizmann.ac.il/content/tns-getting-started
14 Smith et al.
Figure 4. Three examples of triplet images, target (left), reference (middle) and difference (right) of objects that have pasedthe automated filters, are associated with galaxies with known redshifts and are presented to human scanners for final selection.They are confirmed supernovae SN2019shb, SN2019tua and SN2019vsa (respectively from top to bottom), the top one has atensorflow real-bogus factor RB=0.71 and the other two have RB=1.
is MyISAM tables can be ordered by a specified index. One property of HTMs is that spatially nearby triangles are
physically close to each other in the database (because they are numerically close). Hence if a table is ordered by the
HTM index, this reduces lookup time. The main disadvantage of MyISAM is that the whole table is locked when a
single row is updated (forcing many parallel updates to be serialised internally). However, parallel inserts are allowed,
and this property is used by the ingester.
The ATLAS database, which has been running since 2015 December and contains 20 billion detections (5.7 billion
objects) currently occupies about 6.2 TB (as of 2020 March 12). The size will likely double with the introduction of
the southern telescopes in Chile and South Africa. The ATLAS postage stamps currently occupy 12TB (mostly junk
or asteroid data, which is kept for future machine learning), but extraction of the stamps also requires download of
the original input, reference and difference images - which currently accounts for 0.2 TB per day per telescope. Hence
a single telescope database is growing by 0.75TB per year, stamps are growing by 1.5TB per year and image data (if
ATLAS Transient Survey 1507/01/2020, 19)22Candidate 1215800291241557000
Page 1 of 3https://star.pst.qub.ac.uk/sne/atlas4/candidate/1215800291241557000/
ATLAS19zjtCoords:
21+58+00.27 +24+15+57.1(329.50116 +24.26588)Galactic (l,b)
(79.79560,-23.73579)
Realbogus Factors
0.36 (DEW) 1.00 (TF)
Flag Date:
Oct. 31, 2019
Generate AstroNote
IAU Name:
2019tuaSee Also:
AT2019tuaTNS Crossmatch:
2019tuaOffset = 0.07"
Object z = 0.010
Discoverer = ATLAS
Discovered on Oct. 31, 2019,
7U52 a.m.
Number of Detections:
170
Object List:
good
Processing Flags:
moons tns mpc reffinders stamps eph
Spectral Type:
SN II
Context:
SNThe transient is possibly associated with UGC11860/1237680308005765563; an r=16.13 mag galaxy
found in the GLADE/NED/NED_D/SDSS catalogues. It's located 2.29" S, 7.26" E (1.5 Kpc) from the
galaxy centre. A host distance of 40.4 Mpc(z=0.010) implies a m - M = 33.03.
Comments:
2019-10-31 12U51U19 (shubham.srivastav): fast track, astronote needed
Unforced Photometry Lightcurve
58760 58780 58800 58820 58840 58860
20.0
19.0
18.0
17.0
16.0
15.0
14.0
−20 0 20 40 60
cocot
mjd
days since earliest detection
AB M
ag
Raw Unforced Data
Current MJD (vertical line): 58855.80694
Forced Photometry Lightcurve
58760 58780 58800 58820 58840 5886021.0
20.0
19.0
18.0
17.0
16.0
15.0 coco
mjd
AB M
ag
Raw Forced Data
Forced Photometry Flux
58740 58760 58780 58800 58820 58840 58860
−2000.0
−1000.0
0.0
1000.0
2000.0
3000.0 co
mjd
Flux
/ µJ
y
Forced Diffstack Photometry Flux
58400 58500 58600 58700 58800
−1000.0
0.0
1000.0
2000.0
3000.0
cot
mjd
Flux
/ µJ
y
−5 0 5
−5
0
5
1st detection
mean coords (rms = 0.09)
1215759891241558000
1215800061241601800
1215800791241559300
ΔRA / arcsec
ΔD
ec /
arcs
ec
Finders:
WIKISKY
PS1 Stamp
SDSS DR15 Navigate
NED: 2 arcmin radius
LEDA: 10 arcsec radius
SIMBAD: 10 arcsec radius
Galactic Extinction
target (c) mjd: 58787.35079 ref (c) mjd: 58787.35079 diff (c) mjd: 58787.35079
target (c) mjd: 58787.33906 ref (c) mjd: 58787.33906 diff (c) mjd: 58787.33906
target (c) mjd: 58787.33182 ref (c) mjd: 58787.33182 diff (c) mjd: 58787.33182
target (c) mjd: 58787.32842 ref (c) mjd: 58787.32842 diff (c) mjd: 58787.32842
det id ra dec mag dmag x y major minor phi det chin pvr ptr pmv pkn pno pbn pcr pxt psc dup wpflx dflx image group id mjd obs mag5sig atlas metadata id texp filt obj fpRA fpDec3790 329.50121 24.26586 18.723 0.096 1148.85 5945.36 2.36 2.06 128.7 0 0.85 815 15 0 0 0 0 0 0 44 1 467.6 1.9 16328270 58787.3284215 02a58787o0268c 19.73 — 30.0 c SJ327N24 327.16606 23.93753
3791 329.50122 24.26587 18.63 0.139 1148.83 5945.39 2.73 2.07 136.9 1 0.91 819 16 8 0 0 0 0 0 40 -1 463.3 1.9 — 58787.3284215 02a58787o0268c 19.73 — 30.0 c SJ327N24 327.16606 23.93753
3963 329.50115 24.26577 19.442 0.126 1187.05 5982.82 1.8 2.08 97.9 2 1.32 751 11 3 0 0 0 0 0 84 -1 482.4 1.9 — 58787.3390604 02a58787o0290c 19.72 — 30.0 c SJ327N24 327.18749 23.91765
3964 329.50114 24.2658 18.813 0.13 1187.06 5982.89 2.41 2.15 109.9 1 1.32 741 10 4 0 0 0 0 0 89 1 492.4 1.9 16328251 58787.3390604 02a58787o0290c 19.72 — 30.0 c SJ327N24 327.18749 23.91765
3965 329.50114 24.26581 18.81 0.08 1187.06 5982.91 2.36 2.08 128.2 0 1.54 738 10 0 0 0 0 0 0 92 -1 493.8 1.9 — 58787.3390604 02a58787o0290c 19.72 — 30.0 c SJ327N24 327.18749 23.91765
4074 329.50123 24.26577 18.648 0.128 1129.41 5905.73 2.89 2.3 133.7 1 1.03 771 14 6 0 0 0 0 0 64 -1 487.9 1.9 — 58787.3318243 02a58787o0275c 19.73 — 30.0 c SJ327N24 327.15511 23.958
4075 329.50126 24.26578 19.69 0.142 1129.37 5905.75 1.19 2.51 144.5 2 0.97 770 13 1 0 0 0 0 0 61 1 475.0 1.9 16328258 58787.3318243 02a58787o0275c 19.73 — 30.0 c SJ327N24 327.15511 23.958
4076 329.50123 24.26579 18.848 0.078 1129.42 5905.76 2.36 2.07 126.4 0 1.52 762 13 0 0 0 0 0 0 74 -1 493.6 1.9 — 58787.3318243 02a58787o0275c 19.73 — 30.0 c SJ327N24 327.15511 23.958
3899 329.50127 24.26583 18.711 0.083 1207.83 5865.25 2.42 2.08 127.0 0 1.46 814 20 0 0 0 0 0 0 43 1 419.2 1.7 16328242 58787.3507883 02a58787o0315c 19.67 — 30.0 c SJ327N24 327.19907 23.97801
3900 329.50125 24.26583 18.446 0.101 1207.86 5865.25 2.91 2.62 121.7 1 2.0 814 20 8 0 0 0 0 0 43 -1 429.1 1.7 — 58787.3507883 02a58787o0315c 19.67 — 30.0 c SJ327N24 327.19907 23.97801
J2000 21 58 0.270 +24 15 57.10
FoV: 1.2'
+–
Figure 5. Candidate object page showing unforced, forced and forced photometry in flux space (µJansky units) lightcurves.(Non-detections in the magnitude plots are shown as arrows, which display the limiting magnitude (5σ for unforced, 3σ forforced.) The most recent four rows of images on the day it was flagged are also displayed, along with a scatter plot of detections(which also shows a passing asteroid). An Aladin-Lite widget is shown displaying Pan-STARRS imagery, and underneath thisare shown two finders from Pan-STARRS data in both PS1 g-band and colour (PS1 gri). Text information on the host galaxyis shown. A crossmatch with the Transient Name Server is also displayed, and the object spectral type is shown to be a type IIsupernova.
16 Smith et al.07/01/2020, 18)52Good Candidates
Page 1 of 2https://star.pst.qub.ac.uk/sne/atlas4/followup_quickview_bs/2/?page=217&nobjects=5
Good Candidates(6620)
Classify
previous Page 217 of 1324. next
All Possible All Eyeball All Attic All Trash
ATLAS19ytu (2019ths) Coords:
00H28H01.32 -04H33H16.9(7.00553 -4.55470)
Galactic Coords
(108.00919,-66.74561)Flag Date:
Oct. 24, 2019Context
SNSpectral Type
SN IInRealbogus Factors
0.06 (DEW) 0.99 (TF)Earliest and Latest MJDs
58778.378 58852.259Aliases
2019ths, AT2019ths
58800 58850
20.0
19.0
18.0
0 50
coco
mjd
days since earliest detection
AB M
ag
Catalogue:
SDSS DR12 PhotoObjAll TableHost:
1237679323922891091
−5 0 5
−5
0
5
1st detection
mean coords (rms = 0.48)
ATLAS19ytv
1002801350043323600
1002801740043316800
1002801260043309700
ΔRA / arcsec
ΔD
ec /
arcs
ec
Possible Attic Trash
ATLAS19ysj (2019tgm) Coords:
22H18H10.96 +45H43H10.0(334.54568 45.71946)
Galactic Coords
(96.91911,-9.27001)Flag Date:
Oct. 24, 2019Context
SNSpectral Type
SN IaRealbogus Factors
0.60 (DEW) 1.00 (TF)Earliest and Latest MJDs
58778.357 58850.233Aliases
2019tgm, AT2019tgm
58780 58800 58820 58840 5886020.0
18.0
16.0
0 20 40 60 80
coco
mjd
days since earliest detection
AB M
ag
Catalogue:
The GLADE Catalogue v2.3Host:
UGC11979z (spec):
0.02Distance:
82.7 MpcDistance Modulus:
34.59Physical Offset:
19.96 Kpc
−5 0 5
−5
0
5
1st detection
mean coords (rms = 0.16)
ΔRA / arcsec
ΔD
ec /
arcs
ec
Possible Attic Trash
ATLAS19ytr (2019thq) Coords:
21H52H48.09 -08H31H37.3(328.20041 -8.52704)
Galactic Coords
(48.12683,-43.92084)Flag Date:
Oct. 24, 2019Context
SNRealbogus Factors
0.05 (DEW) 0.79 (TF)Earliest and Latest MJDs
58632.558 58788.302Aliases
2019thq, AT2019thq
58400 58600 5880020.0
18.0
16.0
14.0
0 200 400
o-cco
mjd
days since earliest detection
AB M
ag
Catalogue:
SDSS DR12 PhotoObjAll TableHost:
1237652599031333251
−5 0 5
−5
0
5
1st detection
mean coords (rms = 1.64)
1215247940083137100
1215248120083143400
1215247680083138400
ΔRA / arcsec
ΔD
ec /
arcs
ec
Possible Attic Trash
All Undecided
Undecided
Undecided
Undecided
00 28 1.320 -04 33 16.90
FoV: 1.19'
22 18 10.960 +45 43 10.0
FoV: 1.19'
21 52 48.090 -08 31 37.30
FoV: 1.19'
Figure 6. A quickview page showing multiple objects in the “good” list. Summary information including lightcurves, scatterplots and Aladin-Lite widgets are presented, along with a representative triplet of postage stamps. Objects can be movedone-at-a time using buttons under the stamps, or all at once using “select all” buttons at the head and foot of the page. Listsof objects can be filtered by (e.g.) context type, or GW event ID, and ordered by (e.g.) RB factor. Note that the galacticcoordinates are highlighted in red if the object is within galactic latitude |b| < 15.
download continues at the current rate) will require 60TB per year per telescope. Several times this space for backups
(twice weekly and monthly) on spinning disk is also reserved.
An additional unsolved problem is the database storage of the photometric time series images on all the individual
reduced images (not the difference images) to provide variable star and AGN lightcurves. Heinze et al. (2018) presentedthe first catalogue of variable stars from ATLAS, which is 4.7 million stars, selected from 142 million stars that each
had 100 photometric points or more. A more ambitious project would be to create an ATLAS billion star database
which would store 1000 points on the lightcurve per year for a billion stars. The size of this database would potentially
be at least 100TB per year, assuming just 100 bytes per detection and the database would need to ingest 64 000
detections per second, assuming 50% utilisation. The current ATLAS Transient Science Server infrastructure is not
yet designed to cope with a database of this size and a different architecture and design would be required to serve all
variables.
Web Interface: A web interface (written in python using the Django Web Framework10, Bootstrap11 and Plotly12)
has been built atop the database to allow the human scanners to visually inspect objects that have a reliable machine
learning score. The web interface is designed to be “responsive” hence will adjust how pages are rendered based on
whether they are being viewed via a mobile or desktop browser. Individual objects are displayed to the user utilising as
much of the space as possible (in desktop orientation) to display the maximum amount of information without having
to excessively scroll down (Figure 5). Users can be presented with lists of objects in either simple tabular format or
10 https://www.djangoproject.com11 https://getbootstrap.com12 https://plot.ly/javascript/
ATLAS Transient Survey 17
Figure 7. Image quality monitoring for ATLAS HKO and MLO for their full history. Cyan points represent measurements inthe c filter and orange in the o filter. Recall that Mauna Loa only has had an o filter in place to the present date. Note that theplots show limits per exposure. A & B: PSF measurements - the recorded FWHM (in arcseconds) measured on the detector.C & D: The 5σ limiting magnitude, which is a function of zeropoint variation (extinction), PSF variation and sky brightness.
with full rendered lightcurves, scatter plots, Aladin-Lite context, summary information and the most recent image
triplet (Figure 6). The latter allows a user to see groups of 50 (by default) objects at a time (adjustable), allowing
them to archive multiple objects or retain for promotion with a single click. The lists may be sorted by (e.g.) RB
factor, latest magnitude and/or filtered to display only subsets of objects (e.g. objects tagged as SN or NT, or objects
associated with a specific LIGO/Virgo event).
5. ON-SKY PERFORMANCE AND MONITORING
As with any survey facility, the sensitivity of an ATLAS image is dependent on atmospheric transparency, sky
background (moon phase) and the image quality or PSF as measured on the detector. We monitor all of these
parameters with measurements made on every on sky science, reduced science frame. The values are are written into
the reduced image FITS header to allow on-sky performance monitoring and visualisation of long term trends. (Most
of this information is also propagated to the .ddc files and hence stored in the database.) Plots of the temporal
variation are given in Figure 7 for the full ATLAS history. Points to note on these plots are:
1. The significant improvement in PSF around MJD 57864 (2017 April 21) is due to the replacement of the Schmidt
correctors on both telescopes. The improvement is initially less obvious on ATLAS-MLO for the reasons stated
below.
2. The throughput at ATLAS-HKO was observed to roll off by 0.2 mag between 58420 (2018 October 29) and 58605
(2019 May 02). This was due to a faulty dehumidifier which emitted an oily residue that gathered on the CCD
window. The window was cleaned on 58606 and immediately the throughput returned to normal. The faulty
dehumidifier has been replaced.
18 Smith et al.
15 16 17 18 19 205 Limit
0
20000
40000
60000
80000
100000
120000
140000
Freq
uenc
y
MLOA
15 16 17 18 19 205 Limit
0
20000
40000
60000
80000
100000
Freq
uenc
y
HKOB
15 16 17 18 19 20Discovery Mag
0
20
40
60
80
100
120
Freq
uenc
y
C
Figure 8. A: Histogram of 5σ limiting magnitudes for MLO in the o band, with the same data as in Figure 7. B: Same as inA, but for HKO with the c and o bands plotted together. C: The discovery magnitude distribution for all ATLAS discoveredtransients reported to the TNS that have been classified as supernovae.
3. The large gap at MJD=58525 (2019 February 11) to 58565 (2019 March 23) was due to an ice storm on Haleakala,
which enforced a long period of dome closure. The storm lasted just a few days but disabled electrical service to
the Haleakala summit for 42 days. The gap is also visible in Figure 1.
4. The ATLAS-MLO sensitivity saw a slow, gradual improvement (due to improving the PSF through focus stability)
until MJD=58715 (2019 August 20). Until that point, the PSF measured on the ATLAS-MLO detector was
significantly worse than on ATLAS-HKO (5.5 arcsec median vs 4.0 arcsec median on HKO). A new STA 1600
detector was installed on MJD=58715 which immediately improved the stellar image widths to a median of 3.8
arcsec, even better than ATLAS-HKO. This was due to charge diffusion on the old CCD which degraded the
FWHM (caused by a manufacturing defect). This effect corresponded to a blurring of approximately 2 pixels
FWHM, in quadrature.
5. The 5σ limiting magnitude displays the expected modulation due to lunar phase and the variation in sky back-
ground.
The histogram of the 5σ limiting magnitudes for each telescope are given in Figure 8 for the calendar year 2019.
This includes all sky conditions and all atmospheric transparencies. As can be seen in both Figure 8 and Figure 7
there is a long tail toward poor quality images with limiting magnitudes m5σ < 16, which we deem not to be of
science quality. These typically make up 2% of the total science images. The median limiting magnitudes for each of
the systems (from 2019) is o < 19.0 mag (ATLAS-HKO) c < 19.6 mag (ATLAS-HKO) and o < 19.0 mag (ATLAS-
MLO). Although ATLAS-MLO has historically had poorer PSF measurements, and a similar zeropoint (i.e. similar
throughout), the fact we employ it in both dark and bright time compensates since ATLAS-HKO is preferentially used
when the moon is above the horizon. Since the new CCD was installed on ATLAS-MLO on MJD=58715, the median
has improved to o < 19.3 mag.
6. SUMMARY OF RESULTS AND DISCOVERIES
As described in Section 4.3, real astrophysical transient sources are selected by the ATLAS Transient Server system,
with the final step being human scanning of all the candidate sources which result from the computational processing
steps. Table 2 indicates that we typically typically have 300-400 objects per day to present for human vetting which
results in a typical mean discovery rate or 10 to 15 real extragalactic transients per day. All these real sources
are automatically registered on the International Astronomical Union’s Transient Name Server (IAU TNS)13, and a
catalogue of our discoveries is easily recovered from the TNS web interface or through simple scripts (e.g. Kulkarni
2020).
To date (2020 March 15), we have registered 5282 transients as primary discoveries, and have detected 10 332 in total
in the “Good” list in our Transient Science Server database (the difference being those discovered by other surveys,
mostly iPTF, ZTF, Gaia, ASASSN and Pan-STARRS). We aim to register extragalactic transients that are not known
AGN on the TNS. Hence, as far as possible we remove objects that are coincident with known AGN and CVs (see
Section 4) and fast declining transients (more than 0.5 mag within the 1 hr quad) that are generally associated with
13 https://wis-tns.weizmann.ac.il/
ATLAS Transient Survey 19
0.0 0.1 0.2 0.3 0.4redshift (z)
0
20
40
60
80
100
120
140
160N
umbe
rZTFATLASASAS-SN
Figure 9. Redshift distribution of all ATLAS discovered and detected extragalactic transients reported to the TNS in 2019that have a spectroscopic redshift. For comparison, the redshift distribution of spectroscopically confirmed supernovae reportedto TNS in 2019 by ZTF and ASAS-SN are also shown. It should be noted that ZTF have dedicated spectroscopic followupfacilities.
faint red objects and are hence M-dwarf flares (Rodrıguez Martınez et al. 2019). While we save these in our database
as real objects, our intention is not to let these slip into the TNS reporting stream.
We also do not report objects that are almost certain CVs, which means a very fast rise of several magnitudes
within 1-3 days (typically a gradient of grad < −0.5 mag day−1) from last non-detection and which have no obvious
host object visible to g, r or i ' 23 mag in the Pan-STARRS 3π archive imaging. While this behaviour is almost
exclusively a sign of CVs and they dominate the sky density of very fast rising transients, it is important not to be
over zealous with this filter, since a number of rare, fast, blue extra-galactic transients inhabit this parameter space. In
addition, if extragalactic transients are more than a projected distance of Rp ' 50 kpc from their host, then association
with the correct host and redshift becomes confusion limited. The discovery of AT2018cow showed how a fast blue
transient with a similar rise as a CV could turn out to be a luminous and unusual transient at 60 Mpc, which radiated
from the hard x-ray to the radio (Smartt et al. 2018; Prentice et al. 2018; Margutti et al. 2019).
The selection of transients from detections on the subtractions (in the .ddc files) to final registration on the TNS
depends on several steps, each of which are not guaranteed 100% completeness. This is also true of other surveys, so a
straight comparison of discoveries does not yield a meaningful result. As a test of completeness of our processing, we
extracted transients discovered by two other surveys during the 6 month period 58620 (2019 May 17) to 58787 (2019
November 16), after the HKO ice-storm and the ATLAS-HKO CCD window fogging issue.
The ASAS-SN project (Holoien et al. 2019) registered 88 objects as discoveries on the TNS above δ = −50◦, and we
searched for these in our data. 13 of these were discovered at RAs with large hour angles, and in an area of sky that
ATLAS had either stopped observing for the season or had yet to start. We could not confirm the existence of 12 of the
88 (and they are as yet unconfirmed on the TNS), suggesting they are not real. This left 53 transients that should be
20 Smith et al.
10h 8h 6h 4h 2h 0h 22h 20h 18h 16h 14h
-75°-60°
-45°-30°
-15°
0°
15°
30°45°
60°75°
co
Figure 10. Hammer projection of the ATLAS transients discovered from the start of the survey until 2020 March 15. Cyan andorange points represent the discovery filter of the object. Green and blue lines are the galactic plane and ecliptic respectively.
in the ATLAS database and registered as objects. All were detected on the ATLAS difference images (multiple times
at S/N > 5), and all 53 objects were in the ATLAS Transient database. Three objects did not make the initial filter
of ≥3, good quality detections, at greater than 5σ, on any one night and two had low machine learning RB scores. A
4% missed detection rate from the machine learning is consistent with the levels we set for selection from the ROC
curve discussed in Section 4.3. One further object was rejected as being associated with a compact galaxy labelled as
stellar in the PS1 catalogue (therefore Sherlock classified it as a VS). Therefore our computer filtering from end to
end appears susceptible to a 10% incompleteness. A further 8 objects were not promoted as real sources by human
scanners, as they judged them not to be reliable astrophysical transients on the first night that they were detected.
This indicates the level of uncertainty introduced when scanners have a small number of detections on which to judge
real or bogus (typically introducing a 15% hit on efficiency). This human scanning incompleteness usually arises from
low significance detections on the first night of detection, and the objects are archived, while subsequent nights will
produce much more reliable and secure detections. (See improvements in section 7 item 3.) A search for ZTF objects
with a peak r < 19 mag and at least 4 detections in Lasair, and a cross-correlation with ATLAS objects showed
similar results. The end to end comparison showed we had approximately 80% efficiency with the losses split roughly
equally between automated computer filtering (basic cuts and machine learning) and human decisions.
In Figure 9 we illustrate the redshift distribution of the spectroscopically classified ATLAS transients in 2019. We
include only the objects that were formally discovered by ATLAS (which we define as first report to the TNS) and
exclude those detected by ATLAS but were discovered by other surveys. (Spectroscopically confirmed objects reported
to TNS by ZTF and ASAS-SN in 2019 are also shown for comparison.) At a redshift of z = 0.15, a discovery magnitude
of m ∼ 19.5 corresponds to M ∼ −19.6 mag, or a typical Ia at peak. Those SNe beyond z = 0.15 are significantly
brighter than the bulk of the normal population. Of the 15 SNe found at z > 0.15, 7 are super-luminous SNe (SLSNe),
whereas the remainder include bright Ia (Ia 91T-like or Ia-CSM), IIn and Ic broad-line events. Figure 11 shows example
light curves of different classes, indicating that in the o-filter, the near-infrared secondary peak in normal type Ia SNe
is clearly detected, allowing SNe Ia to be distinguished from SNe Ib/c. This will assist SN rate calculations for the
supernovae for which we do not have a spectroscopic classification.
6.1. Spectroscopic types within the local 100 Mpc volume
An interesting scientific survey that ATLAS will contribute to is measuring complete, volumetric rates of transients
within the Local Universe. At a distance of 100 Mpc (µ = 35), we are sensitive to transients of absolute magnitude of
M . −16 (in regions of low Milky Way extinction) and is an upper limit to the distance within which we will carry
ATLAS Transient Survey 21
58600 58650 58700 58750 58800MJD + offset
102
103
104
Flux
(Jy
) + o
ffset
SN IaSN IcSN II-pecSN IIP
Figure 11. Well-sampled o-band ATLAS light curves of 4 nearby SNe of types Ia, Ic, IIP and II-peculiar. The “shoulder”or secondary peak in the light curve of the SN Ia ∼ 30 days after the primary maximum is seen clearly, whereas the strippedenvelope SNe (SESNe: Ib, Ic and IIb) do not exhibit this feature. The light curves were shifted along the x and y axes forclarity.
out this survey. It is also approximately the upper distance limit to which we would be sensitive to faint, fast fading
kilonova type objects like AT2017gfo (which peaked at Mr,i = −16 mag, within the first 24 hrs; see the early combined
r and i photometric evolution in Andreoni et al. 2017; Arcavi et al. 2017; Coulter et al. 2017; Soares-Santos et al.
2017; Smartt et al. 2017; Valenti et al. 2017).
A total number of 540 transients, associated with host galaxies within 100 Mpc, were detected by ATLAS and
reported to the TNS between 2017 September 22 (MJD 58018.5) and 2019 December 03 (MJD 58820.2). These
22 Smith et al.
Other
SN Ia
29.6%
SESN
11.5%
SN II
37.4%
Stellar
1.1%0.6%
Unclassified
19.8%
A
SN Ia
37.0%
SESN
14.3%
SN II46.7%
Other + Stellar
2.1%
B
Figure 12. Pie charts of the spectroscopic classifications of ATLAS detected transients associated with galaxies with distanceswithin 100 Mpc. A: All ATLAS transients. B: Subset of spectroscopically classified transients. SESN represents the family ofstripped envelope SNe that includes types Ib/c, IIb and Ic broad-line. Stellar transients include Galactic CVs and variable starsas well as extragalactic novae, whereas the “Other” category includes a TDE, ILRTs and LBVs. A total of 540 transients wereincluded, out of which 107 transients are without a spectroscopic classification.
galaxies had either distances or spectroscopic redshifts (in our three primary catalogues: NED, GLADE and SDSS)
and the association radius was set at a sky projected distance of 50 kpc in Sherlock. These include transients discovered
by other surveys, but all were independently detected by ATLAS. Out of these, 107 transients (19.8%) do not have a
reported classification on TNS. Of the 433 transients with a reported spectroscopic classification, 6 (1.1%) were stellar
sources, i.e. foreground Galactic CVs and variable stars or extragalactic novae. Tidal disruption events (TDEs),
intermediate luminosity red transients (ILRTs) and luminous blue variable outbursts (LBVs) together constitute a
small fraction (0.6%) of the classified transients. The breakdown of the classifications is shown in Figure 12 and we
emphasise that this should not be taken as a true measure of relative rates since we have carried out no correction for
completeness as a function of absolute magnitude of each type. It is purely illustrative of the numbers of transients
detected which are associated with galaxies within this distance limit and shows that about 1 object per day is found
which is plausibly within 100 Mpc. As an illustrative comparison we can compare our results to the Lick Observatory
Supernova Search sample (LOSS; Li et al. 2011b,a), which was a fully corrected, volume limited survey of catalogued
galaxies within 60 - 80 Mpc, from a sample of 180 SNe in total. The relative rate for core collapse SNe (CCSNe) in our
sample (433 objects after removing unclassified transients) is 61.0%, whereas SNe Ia account for 37.0%. In comparison,
the relative rate of CCSNe and SNe Ia was 60.6% and 37.7% respectively, in the Lick Observatory Supernova Search
sample (LOSS; Li et al. 2011a).
Within this 100 Mpc volume, we can estimate the simple “observed” volumetric rates for supernovae. Assuming that
ATLAS covers the whole sky down to δ ∼ −50◦ (88% of 4π) and that ±60 degrees in RA from the sun is unobservable
and lost (33% of 4π), then ATLAS covers approximately 60% of the full 4π sky with a 2 day cadence. The number
of type Ia and CCSNe that we have detected within 100 Mpc implies that the lower limit to the rate of supernova is
RIa = 5.3 × 104 Gpc−3 yr−1 for SNe Ia and RCC = 9.0 × 104 Gpc−3 yr−1 for CCSNe. This is based on all ATLAS
detections, not just on discoveries and these volumetric rates represent a hard lower limit, since correction factors
involving sensitivity and sky coverage have not been taken into account. When we calculate such corrections they
will only have the effect of significantly increasing the rates since we are not 100% efficient in detecting SNe within
this volume. The above is based on assuming the 540 transients contain 2% of non-supernovae (which we know is the
fraction of the 433 classified transients), and that the rest are supernovae with a relative rate the same as in the 433
classified events. One may worry about a bias introduced by the selection of transients for spectroscopic follow-up.
False positives such as mistaken CVs, outbursting stars or non-astrophysical artefacts would lie in this un-classified
list simply because those with spectroscopic resources considered them not likely to be real supernovae. We have
visually inspected all lightcurves of these unclassified 107 objects. There are 6 which are likely either CVs, stars or
ATLAS Transient Survey 23
poor subtractions leading to false positives. All the other 101 have clear supernova like lightcurves and are associated
with 100 Mpc galaxies with no clear candidate for a background galaxy observed to Pan-STARRS1 3π depths. Many
of the objects are toward the fainter end of the distribution, m & 18.5 mag and some have incomplete lightcurves due
to seasonal and or weather constraints. However they are all very likely SNe of some type and a future paper will
address classification of these from their lightcurves, with a particular focus on the “shoulder” feature in the o-band
lightcurve which distinguishes type Ia SNe from the others. The 6 possible impostor objects do not influence the above
calculation.
It is interesting to note that these observed rates, with no corrections are already higher than those previously
published at low redshift. For type Ia SNe, Li et al. (2011a) report a rate of 3±0.6×104 Gpc−3 yr−1 within 80 Mpc for
targeted galaxy searches and Frohmaier et al. (2019) report ≈ 2− 3× 104 Gpc−3 yr−1 within z . 0.09 (400 Mpc). For
CCSNe, LOSS find a rate of 7.1±1.6×104 Gpc−3 yr−1 Li et al. (2011a), which is higher (but not seriously discrepant,
considering the uncertainties) to that from Cappellaro et al. (1999) of 4±2×104 Gpc−3 yr−1. Our estimate of the type
Ia rate is nearly a factor 2 higher than currently accepted, which is quite a difference and the discrepancy would only
increase when a proper efficiency and survey simulation calculation is carried out. The CCSN rate is less discrepant, at
30% higher than the LOSS Li et al. (2011a) value. However the completeness correction for CCNSe is likely to be much
larger than for type Ia, due to their fainter luminosities (Li et al. 2011b,a), and we are certainly far from complete
for a Mo ' −16 mag core-collapse SN at 100 Mpc (since, o = µ + Mo + Ao & 19 mag). Therefore it is possible that
some of the SNe which are classified as type Ia on the TNS are actually type Ic (or Ib), since this mis-identification is
often made (Clocchiatti et al. 2000). This would decrease the Ia rate and increase the CCSN rate, although the total
SN rate would be preserved, and that is already 50% higher than the published values. The ATLAS SN counts give
a total SN rate of 1.4 × 105 yr−1 Gpc−1, compared to the LOSS total of 1.0 × 105 yr−1 Gpc−1. Another possibility
is that both the type Ia and CCSN rate really are significantly higher than measured in LOSS and indeed all other
low redshift surveys. This could be due to a higher specific SN rate in lower mass galaxies, meaning targeted galaxy
surveys will be biased against the high rates of production in low mass galaxies. This type of effect has been suggested
by ASAS-SN for type Ia SNe (see the ASASSN results in an all-sky survey Brown et al. 2019). If the core-collapse SN
rate is significantly higher than the LOSS rates, and if such a systematic underestimate were also present at higher
redshifts, then it may close the discrepancy between the cosmic core-collapse rate and the star formation rate from
other indicates as noted by Horiuchi et al. (2011). Although Botticella et al. (2012) find that very local SN rates and
the traditional star-formation rate indicators (Hα and FUV fluxes) are discrepant in the opposite sense. To address
these important problems, we will undertake further careful calculation of the local SN rates in future papers. At face
value, these simple estimates indicate that the accepted local Universe supernova rates could be too low by a factor
1.5 to 2.
7. CONCLUSIONS AND FUTURE IMPROVEMENTS
The ATLAS system now routinely surveys the whole sky visible from Hawaii (above δ > −50◦) every two nights
with the twin telescope system. This results in 10-15 extragalactic transients discovered or detected (where the latter
means discovered first by other surveys) every night with a 5σ detection limit of o < 19 mag. All discoveries are
made public promptly, however the spectroscopic classification completeness is only about 25%. The processing at
QUB, during day time while the ATLAS telescopes are observing produces rapid discoveries now routinely during the
Hawaiian night. Computer algorithms for filtering detections focus on real-bogus classifications, removal of moving
objects and stars, and a combination of boosted decision trees and machine learning provides an efficient method to
present users with a short list of real extragalactic transients. However human vetting is still essential in this last step
to ensure an optimal balance between false positive rate and missed detection rate. The ATLAS discoveries not only
enable rapid follow-up of interesting sources (e.g. AT2018cow Prentice et al. 2018) but will provide complete statistical
samples within the local volume of 100 Mpc.
Looking forward, we foresee the following improvements to the system
1. Addition of two more units in the southern hemisphere at the South African Astronomical Observatory, and El
Sauce in Chile. These will allow the full sky to be covered every night with our quad sequence and transients to
be discovered routinely within 12-24 hrs. An assessment of the data flow speed to QUB and what extra hardware
is required is essential. It will be prudent to review if the QUB Transient Science Server should continue to be
the main processing node or if Chilean/South African resources (such as the ALeRCE architecture) can improve
the speed and reliability and if there should be multiple sources of data release for the transients.
24 Smith et al.
2. Process the co-added 4× 30 sec exposures each night as one, deeper 120 sec exposure. We now make a co-added
nightly stack of the difference images, which results in a single image deeper by 0.75 magnitudes. Implementing
our 5σ detection process on these images should enable us to reach o < 20 mag and c < 20.3 mag routinely. We
are experimenting with the real-bogus balance and how to robustly select objects with one detection on nightly
stack. During image subtraction, it would be worthwhile to pursue improvements to the convolution calculation,
optimised for our sampling and point-spread-function. This should help decrease the false positive rate on the
co-added difference images.
3. Development of code that will resurrect any object that has been rejected by human based on low signal-to-noise
detections if new detections arise that are more significant.
4. Application of a machine learning algorithms that will classify objects primarily by their lightcurve (e.g. RAPID
Muthukrishna et al. 2019). However combining this with host object information (galaxy redshift, offset from
the galaxy centre, morphology, and colour) should lead to improvements in the prediction of supernova/transient
type from an incomplete lightcurve.
5. Rapid release of transients as Kafka alerts, as is currently done with ZTF, and combining automatically with
ZTF to provide extra time resolution.
6. Improvements to the input galaxy catalogue, with more spectroscopic redshifts and any reliable photometric
redshift measurements (e.g. Takada et al. 2014; DESI Collaboration et al. 2016; de Jong et al. 2019).
7. Further, closer ties to follow-up programmes to increase spectroscopic classification rates and early data.
8. Creation of a public user interface that provides access to all ATLAS lightcurve data and all information that
is currently proprietary to the science teams. This is a significant effort to provide both the computing and
software resources to support a large, public user base.
9. Creation of a large database of variability of every object in the sky brigther than o ∼ 19 mag. And to provide
public access as in the previous point.
8. ACKNOWLEDGEMENTS
This work has made use of data from the Asteroid Terrestrial-impact Last Alert System (ATLAS) project. ATLAS
is primarily funded to search for near earth asteroids through NASA grants NN12AR55G, 80NSSC18K0284, and
80NSSC18K1575 (under the guidance of Lindley Johnson and Kelly Fast); byproducts of the NEO search include images
and catalogs from the survey area. The ATLAS science products have been made possible through the contributions of
the University of Hawaii Institute for Astronomy, the Queen’s University Belfast, the Space Telescope Science Institute,
and the South African Astronomical Observatory. We acknowledge support for transient science exploitation from the
EU FP7/2007-2013 ERC Grant agreement no [291222], STFC Grants ST/P000312/1, ST/N002520/1, ST/S006109/1
and support from the QUB Kelvin HPC cluster, and the QUB International Engagement Fund. TWC acknowledges
EU Funding under Marie Sk lodowska-Curie grant agreement No 842471.
REFERENCES
Abadi, M., Agarwal, A., Barham, P., et al. 2015,
TensorFlow: Large-Scale Machine Learning on
Heterogeneous Systems, software available from
tensorflow.org
Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2017a,
PhRvL, 119, 161101
Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2017b,
ApJL, 848, L12
Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2017c,
ApJ, 848, L13
Alam, S., Albareti, F. D., Allende Prieto, C., et al. 2015,
The Astrophysical Journal Supplement Series, 219, 12
Alard, C. & Lupton, R. H. 1998, ApJ, 503, 325
Andreoni, I., Ackley, K., Cooke, J., et al. 2017, PASA, 34,
e069
Arcavi, I., Hosseinzadeh, G., Howell, D. A., et al. 2017,
Nature, 551, 64
Astier, P., Guy, J., Regnault, N., et al. 2006, A&A, 447, 31
Astropy Collaboration, Robitaille, T. P., Tollerud, E. J.,
et al. 2013, A&A, 558, A33
ATLAS Transient Survey 25
Baltay, C., Rabinowitz, D., Hadjiyska, E., et al. 2013,
PASP, 125, 683
Becker, A. 2015, HOTPANTS: High Order Transform of
PSF ANd Template Subtraction, Astrophysics Source
Code Library
Bellm, E. C., Kulkarni, S. R., Graham, M. J., et al. 2019,
PASP, 131, 018002
Blagorodnova, N., Neill, J. D., Walters, R., et al. 2018,
PASP, 130, 035003
Boch, T. & Fernique, P. 2014, Astronomical Society of the
Pacific Conference Series, Vol. 485, Aladin Lite: Embed
your Sky in the Browser, ed. N. Manset & P. Forshay, 277
Bonnarel, F., Fernique, P., Bienayme, O., et al. 2000,
A&AS, 143, 33
Botticella, M. T., Smartt, S. J., Kennicutt, R. C., et al.
2012, A&A, 537, A132
Brink, H., Richards, J. W., Poznanski, D., et al. 2013,
MNRAS, 435, 1047
Brown, J. S., Stanek, K. Z., Holoien, T. W. S., et al. 2019,
MNRAS, 484, 3785
Brown, P. J., Holland, S. T., Immler, S., et al. 2009, AJ,
137, 4517
Bufano, F., Immler, S., Turatto, M., et al. 2009, ApJ, 700,
1456
Cabrera-Vives, G., Reyes, I., Forster, F., Estevez, P. A., &
Maureira, J.-C. 2017, ApJ, 836, 97
Cappellaro, E., Evans, R., & Turatto, M. 1999, A&A, 351,
459
Chambers, K. C., Magnier, E. A., Metcalfe, N., et al. 2016,
ArXiv e-prints [1612.05560]
Chen, T. W., Inserra, C., Fraser, M., et al. 2018, ApJL,
867, L31
Chollet, F. et al. 2015, Keras,
https://github.com/fchollet/keras
Clocchiatti, A., Phillips, M. M., Suntzeff, N. B., et al. 2000,
ApJ, 529, 661
Coulter, D. A., Foley, R. J., Kilpatrick, C. D., et al. 2017,
Science, 358, 1556
Dalya, G., Galgoczi, G., Dobos, L., et al. 2018, Monthly
Notices of the Royal Astronomical Society, 479, 2374
de Jong, R. S., Agertz, O., Berbel, A. A., et al. 2019, The
Messenger, 175, 3
DESI Collaboration, Aghamousa, A., Aguilar, J., et al.
2016, arXiv e-prints, arXiv:1611.00036
Downes, R. A., Webbink, R. F., Shara, M. M., et al. 2001,
The Publications of the Astronomical Society of the
Pacific, 113, 764
Drake, A. J., Djorgovski, S. G., Mahabal, A., et al. 2009,
ApJ, 696, 870
Dumoulin, V. & Visin, F. 2016, arXiv preprint
arXiv:1603.07285
Fabbri, J., Otsuka, M., Barlow, M. J., et al. 2011, MNRAS,
418, 1285
Flesch, E. W. 2019, arXiv.org, arXiv:1912.05614
Flewelling, H. A., Magnier, E. A., Chambers, K. C., et al.
2016, ArXiv e-prints [1612.05243]
Forster, F. 2019, Transient Name Server Discovery Report,
2019-1403, 1
Frohmaier, C., Sullivan, M., Nugent, P. E., et al. 2019,
MNRAS, 486, 2308
Gaia Collaboration, Brown, A. G. A., Vallenari, A., et al.
2018, A&A, 616, A1
Gaia Collaboration, Brown, A. G. A., Vallenari, A., et al.
2016, A&A, 595, A2
Gal-Yam, A., Arcavi, I., Ofek, E. O., et al. 2014, Nature,
509, 471
Gal-Yam, A., Yaron, O., Sass, A., et al. 2019, Transient
Name Server AstroNote, 1, 1
Goldstein, D. A., D’Andrea, C. B., Fischer, J. A., et al.
2015, AJ, 150, 82
Graham, M. J., Kulkarni, S. R., Bellm, E. C., et al. 2019,
PASP, 131, 078001
Hastie, T., Tibshirani, R., & Friedman, J. 2009, The
elements of statistical learning: data mining, inference,
and prediction (Springer Science & Business Media)
Heinze, A. 2019a, Central Bureau Electronic Telegrams,
4645
Heinze, A. 2019b, Central Bureau Electronic Telegrams,
4648
Heinze, A. N., Tonry, J. L., Denneau, L., et al. 2018, AJ,
156, 241
Holoien, T. W. S., Brown, J. S., Vallely, P. J., et al. 2019,
MNRAS, 484, 1899
Holoien, T. W. S., Stanek, K. Z., Kochanek, C. S., et al.
2017, MNRAS, 464, 2672
Horiuchi, S., Beacom, J. F., Kochanek, C. S., et al. 2011,
ApJ, 738, 154
Howell, D. A., Sullivan, M., Perrett, K., et al. 2005, ApJ,
634, 1190
Immler, S., Brown, P. J., Milne, P., et al. 2006, ApJL, 648,
L119
Immler, S. & Lewin, W. H. G. 2003, X-Ray Supernovae, ed.
K. Weiler, Vol. 598, 91–111
Inserra, C., Smartt, S. J., Jerkstrand, A., et al. 2013, ApJ,
770, 128
Kasen, D. & Bildsten, L. 2010, ApJ, 717, 245
Kingma, D. P. & Ba, J. 2014, arXiv preprint
arXiv:1412.6980
26 Smith et al.
Kleyna, J. T., Hainaut, O. R., Meech, K. J., et al. 2019,
The Astrophysical Journal, 874, L20
Kotak, R., Meikle, P., Pozzo, M., et al. 2006, ApJL, 651,
L117
Kromer, M. & Sim, S. A. 2009, MNRAS, 398, 1809
Kulkarni, S. R. 2020, Transient Name Server AstroNote, 1,
1
Lasker, B. M., Lattanzi, M. G., McLean, B. J., et al. 2008,
The Astronomical Journal, 136, 735
Law, N. M., Kulkarni, S. R., Dekany, R. G., et al. 2009,
PASP, 121, 1395
Leaman, J., Li, W., Chornock, R., & Filippenko, A. V.
2011, MNRAS, 412, 1419
Li, W., Chornock, R., Leaman, J., et al. 2011a, MNRAS,
412, 1473
Li, W., Leaman, J., Chornock, R., et al. 2011b, MNRAS,
412, 1441
Lochner, M., McEwen, J. D., Peiris, H. V., Lahav, O., &
Winter, M. K. 2016, ApJS, 225, 31
Magnier, E. A., Chambers, K. C., Flewelling, H. A., et al.
2016a, ArXiv e-prints [1612.05240]
Magnier, E. A., Schlafly, E. F., Finkbeiner, D. P., et al.
2016b, ArXiv e-prints [1612.05242]
Magnier, E. A., Sweeney, W. E., Chambers, K. C., et al.
2016c, ArXiv e-prints [1612.05244]
Mahabal, A., Rebbapragada, U., Walters, R., et al. 2019,
PASP, 131, 038002
Margutti, R., Kamble, A., Milisavljevic, D., et al. 2017,
ApJ, 835, 140
Margutti, R., Metzger, B. D., Chornock, R., et al. 2019,
ApJ, 872, 18
McBrien, O. R., Smartt, S. J., Chen, T.-W., et al. 2019,
ApJL, 885, L23
Muthukrishna, D., Narayan, G., Mandel, K. S., Biswas, R.,
& Hlozek, R. 2019, PASP, 131, 118002
Narayan, G., Zaidi, T., Soraisam, M. D., et al. 2018, ApJS,
236, 9
Patterson, M. T., Bellm, E. C., Rusholme, B., et al. 2019,
PASP, 131, 018001
Perlmutter, S., Gabi, S., Goldhaber, G., et al. 1997, ApJ,
483, 565
Pignata, G., Maza, J., Antezana, R., et al. 2009, in
American Institute of Physics Conference Series, Vol.
1111, American Institute of Physics Conference Series,
ed. G. Giobbi, A. Tornambe, G. Raimondo, M. Limongi,
L. A. Antonelli, N. Menci, & E. Brocato, 551–554
Prentice, S. J., Maguire, K., Smartt, S. J., et al. 2018,
ApJL, 865, L3
Price-Whelan, A. M., Sipocz, B. M., Gunther, H. M., et al.
2018, AJ, 156, 123
Quimby, R. M. 2006, PhD thesis, The University of Texas
at Austin
Quimby, R. M., Kulkarni, S. R., Kasliwal, M. M., et al.
2011, Nature, 474, 487
Rabinak, I. & Waxman, E. 2011, ApJ, 728, 63
Reyes, E., Estevez, P. A., Reyes, I., et al. 2018, arXiv
e-prints, arXiv:1808.03626
Ritter, H. & Kolb, U. 2003, Astronomy and Astrophysics,
404, 301
Rodrıguez Martınez, R., Lopez, L. A., Shappee, B. J., et al.
2019, arXiv e-prints, arXiv:1912.05549
Ryder, S. D., Sadler, E. M., Subrahmanyan, R., et al. 2004,
MNRAS, 349, 1093
Schmidt, B. P., Suntzeff, N. B., Phillips, M. M., et al. 1998,
ApJ, 507, 46
Skrutskie, M. F., Cutri, R. M., Stiening, R., et al. 2006, AJ,
131, 1163
Smartt, S. J., Chen, T. W., Jerkstrand, A., et al. 2017,
Nature, 551, 75
Smartt, S. J., Clark, P., Smith, K. W., et al. 2018, The
Astronomer’s Telegram, 11727, 1
Smartt, S. J., Valenti, S., Fraser, M., et al. 2015, A&A, 579,
A40
Smith, K. W., Denneau, L., Vincent, J. B., & Weryk, R.
2019a, Central Bureau Electronic Telegrams, 4594
Smith, K. W., Williams, R. D., Young, D. R., et al. 2019b,
Research Notes of the American Astronomical Society, 3,
26
Soares-Santos, M., Holz, D. E., Annis, J., et al. 2017, ApJL,
848, L16
Soderberg, A. M., Berger, E., Page, K. L., et al. 2008,
Nature, 453, 469
Soderberg, A. M., Chakraborti, S., Pignata, G., et al. 2010,
Nature, 463, 513
Sonnett, S., Meech, K., Jedicke, R., et al. 2013, PASP, 125,
456
Srivastav, S., Smartt, S. J., Leloudas, G., et al. 2020, arXiv
e-prints, arXiv:2001.09722
Stalder, B., Tonry, J., Smartt, S. J., et al. 2017, ApJ, 850,
149
Stockdale, C. J., Williams, C. L., Weiler, K. W., et al. 2007,
ApJ, 671, 689
Sutaria, F. K., Chandra, P., Bhatnagar, S., & Ray, A. 2003,
A&A, 397, 1011
Szalay, A. S., Gray, J., Fekete, G., et al. 2005, Indexing the
Sphere with the Hierarchical Triangular Mesh, Tech.
Rep. MSR-TR-2005-123
Tachibana, Y. & Miller, A. A. 2018, PASP, 130, 128001
Takada, M., Ellis, R. S., Chiba, M., et al. 2014, PASJ, 66,
R1
ATLAS Transient Survey 27
Tomaney, A. B. & Crotts, A. P. S. 1996, AJ, 112, 2872
Tonry, J. L. 2011, PASP, 123, 58
Tonry, J. L., Denneau, L., Heinze, A. N., et al. 2018,
Publications of the Astronomical Society of the Pacific,
130, 064505
Valenti, S., David, Sand, J., et al. 2017, ApJL, 848, L24
Veres, P., Jedicke, R., Wainscoat, R., et al. 2009, Icarus,
203, 472
Veron-Cetty, M. P. & Veron, P. 2010, Astronomy and
Astrophysics, 518, A10
Waters, C. Z., Magnier, E. A., Price, P. A., et al. 2016,
ArXiv e-prints [1612.05245]
Wood-Vasey, W. M., Miknaitis, G., Stubbs, C. W., et al.
2007, ApJ, 666, 694
Wright, D. 2015, PhD thesis, Department of Physics and
Astronomy, Queen’s University Belfast.
Wright, D. E., Lintott, C. J., Smartt, S. J., et al. 2017,
MNRAS, 472, 1315
Wright, D. E., Smartt, S. J., Smith, K. W., et al. 2015,
MNRAS, 449, 451
Yaron, O. & Gal-Yam, A. 2012, PASP, 124, 668
Zackay, B., Ofek, E. O., & Gal-Yam, A. 2016, ApJ, 830, 27